Journal Publication:

• Title: Development of a Korean News Question Answering System Based on LLM Using RAPTOR
• Authors: Yoseob Yun*, Hyunchul Ahn**
• Abstract: This study proposes a question-answering system based on RAPTOR (Recursive Abstractive Processing for Tree-Organized Retrieval) to address limitations of large language models (LLMs), such as a lack of up-to-date information and the hallucination phenomenon. For comparative analysis, RAG (Retrieval-Augmented Generation), widely used in the industry, was selected as a baseline model. Weekly issue data from BIGKINDS, a news analysis platform in South Korea, was collected and embedded into a vector database to build an external knowledge repository. Using LangChain, an experimental environment was established, and empirical analyses were conducted. The results demonstrated that the proposed RAPTOR method outperformed RAG in answering abstract questions, while both techniques exhibited similar performance for specific queries. These findings highlight RAPTOR’s effectiveness in responding to complex questions and its potential as a solution, particularly for handling abstract queries.
• Keywords: RAPTOR, RAG, LLM, LangChain, Question Answering System
• Journal: Journal of Intelligent Information Systems
• Volume and Issue: Vol. 30, No. 4
• Publication Date: December 2024
• Link: Journal of Intelligent Information Systems, Vol. 30, No. 4

Conference Presentation:

• Main Theme: AI Transformation and Trustworthiness: Shaping the Future of Ethical Intelligence
• Title: Development of a Korean News Question Answering System Based on LLM Using RAPTOR
• Session: D4. [Academic Session] LLM Applications (16
  ~17
  )
• Conference: 2024 Fall International Conference of the Korean Intelligent Information Systems Society
• Date: November 1, 2024
• Location: Kyung Hee University

This publication and presentation detail the development of a Korean news question-answering system leveraging RAPTOR to enhance the capabilities of large language models. The research addresses common limitations in LLMs and demonstrates the effectiveness of RAPTOR in handling complex, abstract queries.